Logistic regression models for combining data from multiple signs and symptoms
Jon Deeks with Siew Chan, Petra Macaskill and Les Irwig (Sydney)
Establishing a diagnosis involves obtaining multiple pieces of diagnostic information from history taking and clinical examination together with laboratory and imaging tests. Combining information from multiple tests is complicated due to dependency and interaction between tests. Combination of likelihood ratios for individual pieces of diagnostic information often uses Bayes’ theorem, but this does not account for test dependency. However, there are links between Bayes’ theorem and logistic regression models that do allow dependency and interaction to be estimated and accounted for.
This methodological study has investigated alternative logistic regression based approaches for analysing data from multiple signs and symptoms. Four methods have been identified and are being compared. As well as standard logistic regression approaches attributed to Albert, Spiegelhalter and Knill-Jones, and Knottnerus have been identified. The methods differ in the manner in which they handle dependency and interaction between tests, as well as the ease with which they can be applied and interpreted.
Using the datasets collected by the Clinical Assessment of the Reliability of the Examination (CARE) group for the diagnosis of chronic obstructive airway disease based on signs and symptoms, we are comparing the four approaches with a simple naïve Bayes approach.
